{"id":"157009c5-4d33-42d6-aadf-9b38f26cbced","arxiv_id":"2501.18973","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"GPO-VAE reinterprets a VAE's perturbation parameters as a gene-by-gene causal matrix, trains it with a differential-expression-matching loss, and reports state-of-the-art perturbation prediction plus GRN inference on three Perturb-seq datasets.","lead":"This paper presents GPO-VAE, a variational autoencoder that predicts how a cell's gene expression changes after a gene is perturbed, and that also yields a gene regulatory network by reinterpreting the model's internal parameters as gene-to-gene causal probabilities. It reports the best accuracy among VAE-based models on three single-cell CRISPR perturbation benchmarks, and the inferred network overlaps with published cancer pathways.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The Bernoulli matrix W∈[0,1] cannot encode downregulation, yet Eq. 15 fits signed ΔX; repression is forced to be 'no edge,' so the inferred GRN and its µWD/FOR scores are biased toward activating edges. A signed-weight retraining test would settle whether the explainability claim survives.","rationale":"The reader's weakest_assumption identified the same signed-ΔX versus [0,1]-Bernoulli mismatch, and I agree that it is the most load-bearing concern. The paper reports concrete numerical results and an ablation, and the perturbation-response gains over CRADLE-VAE are plausible as regularization effects. However, the central novelty and explainability contribution rest entirely on interpreting W as a causal gene regulatory network. If W cannot represent repression, then the inferred graph is incomplete in a biologically essential direction, and the statistical GRN evaluation is biased toward edges whose targets are upregulated. This is a correctness risk internal to the model's stated semantics, not a disagreement with external consensus. The proposed synthetic test would decide whether the issue is inherent or merely an implementation detail. Because the flaw is fixable by reparameterizing W with signed weights or an explicit inhibition matrix, and because the perturbation-prediction results may still stand, the appropriate verdict remains conditional; no adjustment to the reader's verdict is needed.","tokens_in":19741,"tokens_out":4136,"duration_ms":45475,"concrete_test":"Train GPO-VAE with the same code and loss on a synthetic perturbation dataset generated from a known GRN containing both activating and repressive edges; measure recall of true repressive edges under the 0.5 threshold. If repressive-edge recall is near zero while activating-edge recall is high, the [0,1] parameterization, not the data, is the binding limitation of the GRN claim.","verdict_should_be":"UNCHANGED","load_bearing_attack":"In §2.2.5, W is explicitly a matrix of causal gene-to-gene probabilities, with scalar values in [0,1] and edges thresholded at 0.5 (§3.2). The GPO loss in Eq. 20, Jgpo(W) = Σ||P T_K − ΔX||_1 + ||W||_1, fits the signed differential expression matrix ΔX directly. The paper never specifies how signed fold changes are mapped into a [0,1] Bernoulli parameter space. Since P T_K is nonnegative under Eq. 18, any downregulated target can only be matched by pushing W toward zero, which is indistinguishable from 'no edge' after thresholding. Consequently, the inferred GRN cannot represent repression, which is a core feature of real gene regulation. This is not a cosmetic issue: the paper's headline biological claim is that W is a causal GRN, and the GRN evaluation metrics (µWD and FOR) measure perturbation-versus-control expression shifts, which include downregulation. With a nonnegative W, selected high-weight edges are systematically biased toward positive-effect targets, so the reported 'best µWD and FOR' partly reflect this representational bias rather than genuine causal recovery. The K-hop term in Eq. 18 compounds the problem, because powers of a nonnegative matrix only accumulate activating paths. A signed reparameterization is needed to test whether the GRN claim survives.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes GPO-VAE, a variational autoencoder for single-cell gene perturbation response prediction in which the latent perturbation effect is governed by a square Bernoulli parameter matrix W that is interpreted as a gene regulatory network (GRN). A GRN-aligned parameter optimization (GPO) loss (Eq. 20) fits W, through a multi-hop accumulated matrix T_K and an L1 sparsity penalty, to the signed differential expression between perturbed and control samples. Experiments on Replogle K562, Replogle RPE1, and Adamson report state-of-the-art ATE correlation/R²/Jaccard among VAE-based baselines, and the GRN obtained by thresholding W at 0.5 is reported to achieve the best µWD and FOR against CausalBench baselines. Qualitative case studies discuss KRAS-, MYC-, and NTRK1-related subnetworks, and a case study reports competitive performance on three held-out perturbation treatments.","tokens_in":19914,"tokens_out":5308,"duration_ms":55422,"significance":"If the central claim holds, the contribution is valuable: a single VAE would simultaneously achieve strong perturbation-response prediction and yield an interpretable, data-consistent regulatory network. The paper ships public code, reports results over multiple seeds with standard deviations, and includes a systematic ablation (Table 5) showing that each GPO loss component contributes. The perturbation-prediction comparisons in Table 3 are internally consistent and support the state-of-the-art claim within the VAE family on these datasets. However, the explainability claim rests on reading W as a matrix of causal probabilities in [0,1], while the GPO loss fits signed differential expression. Because P T_K is nonnegative under the stated parameterization, the model cannot represent downregulation except by pushing W toward zero, which is indistinguishable from 'no edge' after thresholding. This is a load-bearing correctness risk for the GRN-inference and biological-interpretation claims, and it also casts doubt on the reported µWD/FOR improvements.","major_comments":[{"comment":"W is described as a matrix of causal gene-to-gene probabilities with entries in [0,1] (also stated in §3.2), and Eq. (1) samples Bernoulli masks from W. Yet no constraint or reparameterization (e.g., sigmoid) is specified, and the GPO loss J_gpo = Σ||P T_K − ΔX||_1 + ||W||_1 fits the signed differential expression matrix ΔX directly. Since P T_K is nonnegative under Eq. (18), a downregulated target can only be fitted by driving W toward zero; after thresholding at 0.5 this is indistinguishable from 'no edge'. The inferred GRN therefore cannot represent repression, which is a core feature of gene regulation. Please reparameterize W with signed weights (or separate positive/negative edge matrices) and retrain, or otherwise show how the current [0,1] parameterization can represent signed fold changes; at a minimum, report the distribution of signs in ΔX and the model's ability to capture them.","section":"§2.2.5, Eq. (15)-(20)"},{"comment":"The assertion that W^k, normalized by 1/|G°∪G+|, 'contains k-hop causal relationships' is stated without derivation. For a matrix with entries in [0,1], powers are not Bernoulli probability matrices and row sums are not preserved, so the normalization is ad hoc; moreover, powers of a nonnegative matrix only accumulate positive regulatory paths, compounding the sign bias identified above. The paper should either provide a derivation of this multi-hop interpretation or validate it empirically, for example by showing that the k-hop terms improve prediction of both upregulated and downregulated targets rather than only positive effects.","section":"§2.2.5, Eq. (18)"},{"comment":"The GRN evaluation metrics µWD and FOR are computed from perturbation-versus-control expression changes, which include strongly downregulated genes. Since the current W cannot represent negative edges, edges with strong negative target shifts cannot be selected, so the reported 'best µWD and FOR' for GPO-VAE in Table 4 may partly reflect this representational bias rather than genuine recovery of causal regulatory structure. The paper should report the signed edge analysis and, ideally, recompute µWD separately for positive- and negative-effect target genes under a signed reparameterization.","section":"§3.2 and Table 4"},{"comment":"The GRN inference result is partly circular: the GPO loss in Eq. (20) explicitly minimizes ||P T_K − ΔX||_1, where ΔX is the differential expression between perturbed and control samples, and the CausalBench-style evaluation in §3.2 scores an edge A→B by the perturbation-vs-control distribution shift of B when A is perturbed. The model is therefore being scored on a quantity closely tied to the loss it was trained to minimize, unlike the unsupervised causal-discovery baselines. The paper should acknowledge this and provide a nontrivial test of the GRN claim, such as held-out perturbation evaluation, comparison against a supervised baseline fitted to ΔX, or evaluation on independent pathway databases with proper significance testing.","section":"§2.2.5 vs. §3.2"}],"minor_comments":[{"comment":"K is introduced only after Eq. (18) ('K is equally set to 5'); please define K before its first use in Eq. (17).","section":"§2.2.5"},{"comment":"The edge threshold is stated as 0.5, but Figure 3 distinguishes 'initialized parameter (=0.5)' from 'presence of edge (>0.5)'. Please clarify whether the threshold is strict (>0.5) or non-strict (≥0.5), since this changes the number of edges in Table 4.","section":"§3.2 and Figure 3"},{"comment":"The Random 100,000 baseline row reports FOR = -1.000 for all datasets; under the definition of FOR as a proportion of negative edges, negative values are not possible. Please check the baseline computation or the metric definition.","section":"Table 4"},{"comment":"The text says the full GPO-VAE gives a 'slight improvement in perturbation response prediction' over its ablations, but on RPE1 the full model has ATE-ρ 0.6584, which is slightly lower than the J_K_dge ablation's 0.6593. Please qualify the statement or note that the improvement is not uniform across metrics.","section":"Table 5"},{"comment":"The pathway case studies are suggestive, but no enrichment p-values or multiple-testing corrections are reported. Please add adjusted enrichment statistics or explicitly describe these analyses as anecdotal.","section":"§3.5"}],"recommendation":"major_revision","confidential_remarks":"The paper's perturbation-prediction results are credible, and the code release is a plus. The main risk is the interpretability story: the sign-mismatch between a [0,1] Bernoulli W and signed differential expression is not a minor presentation issue but a structural limitation that directly affects the GRN inference claims, including the µWD/FOR results in Table 4. The fix—a signed reparameterization with retraining—is nontrivial but within the scope of a major revision. The included limitation paragraph is candid about CausalBench evaluation, but it does not address the sign issue, which is more fundamental. I also note that the multi-hop interpretation in Eq. (18) is currently an assertion rather than a derivation; it should be justified or empirically supported."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"GPO-VAE is a genuinely new variant of the sparse-mechanism-shift VAE family: it reshapes the perturbation mask parameter into a square gene-by-gene matrix, treats entries as causal probabilities, and adds a DGE-matching loss with K-hop accumulation and sparsity. The perturbation prediction results on Replogle K562/RPE1 and Adamson are solid, the ablation shows each loss component contributes, and the authors publish code and are upfront about limitations like multi-gene support. That is real work and worth a serious referee.\n\nThe soft spots are not minor. The central explainability claim—that W is a causal GRN—has a load-bearing flaw that the stress-test note correctly identifies. W entries are Bernoulli parameters in [0,1], but the loss in Eq. 20 fits signed differential expression ΔX. Downregulated targets can only be represented by pushing W to zero, which is indistinguishable from no edge after the 0.5 threshold. So the inferred GRN cannot encode repression, and the K-hop powers of a nonnegative matrix only accumulate activating paths. The paper never says how the [0,1] range is enforced or how negative fold changes enter the loss; that is a genuine gap.\n\nThe GRN evaluation is also more circular than the paper admits. µWD and FOR measure perturbation-versus-control expression shifts, which is exactly what the GPO loss fits. The strong µWD/FOR numbers are therefore largely a consistency check, not independent causal verification. The authors do note the biological/statistical mismatch, but they don't acknowledge that the statistical GRN metric is trained into the model.\n\nLesser issues: no sensitivity analysis for the hop count K or the GPO weight β, and the 'state-of-the-art' claim is only against other VAEs. Non-VAE methods like GEARS or simple linear baselines are not compared, so the prediction claim is narrower than the abstract implies.\n\nThe biological case studies are post-hoc and selective, but they are at least external anchors and well-cited.\n\nVerdict: this deserves peer review, not desk rejection. The prediction results and the GPO mechanism are worth publishing after major revision. The authors should either allow signed weights (e.g., separate activation/repression parameters) or explicitly restrict the GRN claim to activating edges, reframe the GRN evaluation as a consistency check, add sensitivity analysis, and benchmark against non-VAE predictors. If they fix the signed issue and the circularity framing, the paper becomes a useful contribution to explainable perturbation modeling.","headline":"A novel VAE with a GRN-shaped mask gives solid perturbation predictions, but the GRN explainability claim is undermined by a signed-expression representation gap and circular evaluation.","tokens_in":20651,"tokens_out":3452,"would_cite":true,"duration_ms":32259,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"GPO-VAE claims that a single variational autoencoder can simultaneously predict transcriptional responses to gene perturbations and recover an interpretable gene regulatory network from the Bernoulli parameter matrix $W$ that generates…","keywords":["gene perturbation response","gene regulatory network","explainable variational autoencoder","sparse mechanism shift","Perturb-seq","causal graph inference","differential expression","K-hop accumulation"],"falsifier":"A direct check: take a known repressor perturbation whose knockdown raises a target's expression; if the model must keep $W$ in $[0,1]$ and the loss only matches $P W$ to the negative differential expression, the predicted effect on that target cannot go negative, so the $W$-as-GRN interpretation fails whenever such repression is observed.","tokens_in":19334,"feed_emoji":"🧬","tokens_out":7191,"duration_ms":68265,"temperature":0.7,"pith_summary":"GPO-VAE is a variational autoencoder for single-cell gene perturbation data that tries to make its latent perturbation effect explicitly interpretable as a gene regulatory network (GRN). The paper's claim is that the same model can predict post-perturbation transcriptional responses at the top of its baseline comparisons and, from the very matrix $W$ that controls its sparse latent perturbation masks, recover a directed gene-to-gene causal graph that agrees with known biology, including KRAS, MYC, and NTRK1 subnetworks, and with statistical tests of causal effect. The mechanism is a GRN-aligned parameter optimization: a differential-expression loss pushes $W$ to reproduce observed fold changes, a $K$-hop accumulation term lets extended genes enter through powers of $W$, and an $ℓ^1$ penalty keeps the graph sparse. If correct, this dissolves the usual trade-off between prediction performance and explainability in perturbation modeling, because the explainable structure is not a post-hoc analysis but the parameterization of the predictor itself.","feed_headline":"One VAE predicts gene responses and recovers the regulatory graph","feed_subtitle":"GPO-VAE turns its latent perturbation masks into a causal gene network, beating dedicated GRN inference baselines.","key_machinery":"The load-bearing object is the square Bernoulli parameter matrix $W \\in [0,1]^{(|G^\\circ \\cup G^+|) \\times (|G^\\circ \\cup G^+|)}$, reinterpreted as a weighted adjacency matrix of gene-to-gene causal probabilities with diagonal entries as self-regulation. It is trained by the GPO objective $J_{\\mathrm{gpo}}(W)=\\|P T_K - \\Delta X\\|_1 + \\|W\\|_1$, where $\\Delta X$ is an optimal-transport-paired differential expression reference, $T_K = W + \\sum_{k=2}^K \\frac{1}{|G^\\circ \\cup G^+|} W^k$ accumulates multi-hop causal paths with $K=5$, and the $ℓ^1$ term enforces sparsity. The same $W$ parameterizes the Bernoulli sampler of the latent perturbation encoder, so the GRN is not a separate inference module but the very mechanism by which the VAE generates perturbation-specific latent offsets, and it is thresholded at 0.5 to produce the reported graph.","core_discovery":"The central discovery is that redesigning the latent perturbation encoder so its Bernoulli mask probabilities form a square matrix $W$ over perturbed and extended genes, and then optimizing $W$ with the GPO objective $J_{\\mathrm{gpo}}(W)=\\|P T_K - \\Delta X\\|_1 + \\|W\\|_1$ where $T_K = W + \\sum_{k=2}^K \\frac{1}{|G^\\circ \\cup G^+|} W^k$, yields both stronger perturbation-response prediction than prior VAE baselines on the Replogle K562, Replogle RPE1, and Adamson datasets, and a sparse thresholded GRN with lower false omission rate and higher mean Wasserstein distance than dedicated causal-discovery baselines. The paper reads each $W_{i,j}$ as the causal probability from gene $i$ to gene $j$, treats $W^k$ as accumulated $k$-hop causal relations, and validates the recovered subnetworks against experimentally supported KRAS, MYC, and NTRK1 pathways. The claim is that one learned object carries both jobs: sampling masks for the VAE and forming a biologically meaningful adjacency graph.","pith_inferences":["Editorial inference: the paper never shows how signed, negative differential expression is encoded by $W \\in [0,1]$, so a direct test of the causal reading is to perturb a known repressor and check whether predicted expression of its targets moves below control levels.","Editorial inference: the term $W^k$ counts walks that may revisit nodes, so the 'k-hop causal relationship' interpretation presumes cycles are meaningful; a DAG-constrained or self-loop-free variant of the GPO loss would clarify whether the multi-hop term accumulates genuine causal paths or mostly adds correlation.","Editorial inference: because the differential expression reference is built from the same training data that measures GRN quality, the reported µWD and FOR improvements could partly reflect fitting to the evaluation signal; a stronger test would be to infer edges for a cell type or perturbation set never used in training and compare against a known reference network.","Editorial inference: if the causal reading of $W$ is correct, the method offers a template for other structured latent spaces where the VAE's sampling distribution is itself the biological quantity of interest, but the probability interpretation needs independent identifiability or external validation before that template can be relied on."],"forward_implications":["If $W$ is a true GRN, every predicted perturbation response is traceable: the change for a target gene is explained by the edges and multi-hop paths through which the perturbed gene influences it.","Because $T_K$ includes paths through extended genes, genes that were never experimentally perturbed can enter explanations, and the model can generalize to unseen perturbation treatments; the paper demonstrates this on held-out TWISTNB, RPL26, and RPL34 perturbations.","The sparse graph recovered from $W$ achieves better mean Wasserstein distance and false omission rate than dedicated GRN inference baselines, so network inference and perturbation prediction are accomplished in a single training run.","Biologically, subnetworks that pass through extended genes nominate candidate interactions, such as KRAS synthetic-lethal partners and MYC- or NTRK1-pathway genes, that curated interaction databases do not contain, yielding concrete testable hypotheses.","The model's architecture does not restrict it to single-gene perturbations, so the same GPO objective could in principle be extended to multi-gene treatments and to synergy or inhibition relationships, although the paper does not implement that extension."],"supporting_citations":[{"why":"provides the Replogle K562 and RPE1 Perturb-seq datasets used for training and evaluation.","marker":"[20]"},{"why":"provides the Adamson Perturb-seq dataset used for training and evaluation.","marker":"[2]"},{"why":"supplies the GRN inference evaluation protocol and the µWD and FOR metrics.","marker":"[4]"},{"why":"contributes the artifact-disentanglement VAE architecture and QC criteria that GPO-VAE extends.","marker":"[5]"},{"why":"introduces the sparse additive mechanism shift VAE, the modeling framework and baseline that GPO-VAE builds on.","marker":"[8]"},{"why":"provides the sparse mechanism shift baseline against which perturbation prediction is compared.","marker":"[10]"},{"why":"supplies the continuous optimization and L1 sparsity penalty idea used in the GPO objective.","marker":"[14]"},{"why":"defines the optimal transport used to pair control and perturbed profiles for the differential expression reference.","marker":"[50]"}],"fun_headline_variants":["Gene perturbation VAE learns causal network from its own latent space","Latent matrix as GRN: one objective predicts and explains","Predicting responses and inferring regulation with one VAE","One VAE: predict responses, extract GRN"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that each entry of the learned matrix $W$, kept between 0 and 1 and matched to signed differential expression through an $ℓ^1$ loss, is a genuine causal gene-to-gene probability, even though a downregulated interaction would have to be encoded as $W$ near 0 and is then indistinguishable from 'no edge' after the 0.5 threshold.","fun_headline_variants_meta":{"raw":{"variants":["Gene perturbation VAE learns causal network from its own latent space","Latent matrix as GRN: one objective predicts and explains","Predicting responses and inferring regulation with one VAE","One VAE: predict responses, extract GRN"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001305,"raw_usage":{"total_tokens":5373,"prompt_tokens":1045,"completion_tokens":4328,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":661,"completion_tokens_details":{"reasoning_tokens":4272}},"tokens_in":661,"tokens_out":4328,"duration_ms":27503,"temperature":1.0,"reasoning_tokens":4272,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-09T21:48:18.445173+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A direct check: take a known repressor perturbation whose knockdown raises a target's expression; if the model must keep $W$ in $[0,1]$ and the loss only matches $P W$ to the negative differential expression, the predicted effect on that target cannot go negative, so the $W$-as-GRN interpretation fails whenever such repression is observed.","supporting_citations":[{"cited_title":"& Others Mapping information-rich genotype-phenotype landscapes with genome-scale Perturb-seq","cited_arxiv_id":null,"evidence_quote":"provides the Replogle K562 and RPE1 Perturb-seq datasets used for training and evaluation."},{"cited_title":"& Others A multiplexed single-cell CRISPR screening platform enables systematic dissection of the unfolded protein response","cited_arxiv_id":null,"evidence_quote":"provides the Adamson Perturb-seq dataset used for training and evaluation."},{"cited_title":"& Karaletsos, T","cited_arxiv_id":null,"evidence_quote":"introduces the sparse additive mechanism shift VAE, the modeling framework and baseline that GPO-VAE builds on."},{"cited_title":"& Regev, A","cited_arxiv_id":null,"evidence_quote":"provides the sparse mechanism shift baseline against which perturbation prediction is compared."},{"cited_title":"& Xing, E","cited_arxiv_id":null,"evidence_quote":"supplies the continuous optimization and L1 sparsity penalty idea used in the GPO objective."},{"cited_title":"& Others Optimal transport: old and new","cited_arxiv_id":null,"evidence_quote":"defines the optimal transport used to pair control and perturbed profiles for the differential expression reference."}],"review_version":1}